Claude Code may be burning your limits with invisible tokens
claude
| Source: HN | Original article
Anthropic’s Claude Code, the company’s AI‑powered coding assistant that many developers have adopted for everything from autocomplete to autonomous bug‑fix loops, is reportedly draining user quotas far faster than advertised. A wave of complaints surfaced in early April, prompting Anthropic to confirm that two independent bugs in the standalone Claude Code binary are inflating token counts by roughly ten‑to‑twenty times. The first bug breaks prompt‑caching, causing the model to resend previously cached context on every iteration. The second bug miscounts tokens generated by the internal “scratchpad” used for chain‑of‑thought reasoning, adding invisible tokens that never appear in the user’s prompt or output but still count against the quota.
The issue matters because Claude Code’s pricing is tied directly to token consumption; developers on the Claude Max plan have seen daily limits evaporate after a single coding session, forcing them to downgrade or switch to competing tools such as OpenAI’s Codex. For teams that rely on autonomous loops to refactor large codebases, the hidden cost threatens project budgets and erodes confidence in Anthropic’s metering transparency. The problem also highlights a broader risk for AI‑augmented development: when internal mechanisms silently consume resources, users cannot predict expenses or optimise prompts.
Anthropic says it is investigating the bugs and will release a hot‑fix within weeks. In the meantime, power users are rolling back to earlier binary versions and disabling the scratchpad feature to curb token burn. Watch for an official patch timeline, any adjustments to quota‑reset policies, and whether Anthropic will introduce more granular token‑usage dashboards. The episode could also spur tighter industry standards for token accounting in AI coding agents, a development that would benefit the Nordic tech ecosystem that increasingly leans on such tools.
Sources
Back to AIPULSEN